Safety_Tuning-Qwen3.5-4B-SFT-STAR-lr5e-5-E4-step240
SFT safety-tuning checkpoint of Qwen/Qwen3.5-4B, exported to Hugging Face safetensors.
Learning rate: 5e-5. Completed optimizer updates: 240. Epoch: 4/4. Global batch size: 128.
SFT uses 7,655 complete teacher trajectories from STAR-1 and STAR-Benign, with reasoning, answer and EOS supervised. Each learning-rate run starts independently from the original base model. No teacher Spec is added to the student context.
Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Source run: qwen35-star-slime-sft-4ep-b128-pro6000-gpu01-20261106-v4.
Native checkpoint iteration: 239 (zero based).
The repository contains inference weights, configuration and tokenizer/processor assets.
Optimizer, RNG state, training logs and training examples are excluded.
All trained text tensors are exported from the native checkpoint. Vision and untrained
MTP tensors are retained from the pinned original base model; MTP was not trained.
Conversion verification includes exact inverse conversion of every native text tensor,
finite-weight checks and matching all 738 Hugging Face tensor keys and shapes.
No evaluation scores are asserted by this model card.